Devin Desktop integrates with a variety of large language models (LLMs), enabling you to tailor AI behavior to your coding style and project requirements. Both Cascade and the Devin Local agent let you choose a model and switch between models at any time; how you do so differs between the two.

Select a language model

Cascade

Devin Desktop provides access to multiple LLMs. To switch between models, use the dropdown list located below the prompt input field in the Cascade interface.

Evaluate several models to determine which best aligns with your development style and task complexity.

Devin Local

Devin Local supports the latest models from Anthropic, OpenAI, Google, and Cognition, as well as several leading open-source models. You can set or switch the model in any of the following ways:
  • Slash command—Run /model with no argument to open the model selector, or switch directly with /model <name>, for example /model opus or /model sonnet.
  • Command flag—Start a session with a specific model: devin --model sonnet -- your prompt.
  • Config file—Set a default model in your Devin config file (~/.config/devin/config.json, or %APPDATA%\devin\config.json on Windows):
    {
  "agent": {
    "model": "swe-1-6-fast"
  }
}

Configure mode

Cascade

After you have selected the LLM, make sure to configure the mode. Cascade offers two distinct operational modes:
  • Code—Enables Cascade to generate and modify code directly within your codebase.
  • Chat—Designed for conversational interactions, this mode supports inquiries about your codebase, programming concepts, and general coding principles.
Note: Always ensure that the Code mode is selected when performing code generation or modification tasks.

Devin Local

Devin Local uses agent modes rather than a Code/Chat toggle:
  • Normal—The default mode. Auto-approves read-only actions and prompts for approval before writing files or running commands. Use this mode for code generation or modification tasks.
  • Plan—Read-only research mode. The agent investigates your codebase and writes up an approach for your approval before making any changes, without modifying code.
  • Ask—For one-off questions about your codebase or ABL concepts, without making code changes.
Switch modes with /normal, /plan, or /ask <question>, or run /mode <name>. Devin Local also offers additional permission modes (Accept Edits, Smart, Bypass, and Autonomous) that control how much approval is required for file edits and shell commands — see your Devin Local permissions documentation for details.
Note: Always ensure that Normal, Accept Edits, or Smart mode is selected when performing code generation or modification tasks; use Plan or Ask only for research and read-only questions.